STG-LSTM: Spatial-temporal graph-based long short-term memory for vehicle trajectory prediction

Abstract

Vehicle trajectory prediction plays a crucial role in enhancing the safety, efficiency, and effectiveness of intelligent transportation systems. Accurate predictions of future vehicle movements are essential for applications such as autonomous driving, traffic management, and collision avoidance systems. However, many existing methods either focus solely on spatial or temporal dimensions, neglecting the dynamic interactions between vehicles, which reduces prediction accuracy, especially in complex traffic scenarios. To address these limitations, the study proposes a Spatial-Temporal Graph-Based Long Short-Term Memory model, which integrates graph convolutional networks with long short-term memory networks to effectively capture both spatial relationships and temporal dependencies in vehicle trajectories. The proposed model employs a proximity-based method to construct dynamic adjacency matrices that represent real-time vehicle interactions. To capture spatial dependencies between vehicles, the study uses graph convolutional networks to model the relationships between neighboring vehicles. The long short-term memory network is then applied to capture temporal dynamics by learning the sequential dependencies in vehicle movement patterns. The output from the long short-term memory network is passed through a fully connected layer, which generates trajectory predictions for each vehicle. The study experimental results demonstrate that the proposed model outperforms existing state-of-the-art models across various prediction metrics. Specifically, at 3s and 4s prediction horizons, the model reduces the root mean square error by 22.4 % and 25.5 %, respectively, compared to the best performing interaction-aware long short-term memory model. At the 5s prediction horizon, the model achieves a significant root mean square error reduction of 26.6 %. These findings highlight the model's potential to improve safety and decision-making in autonomous driving systems and traffic management applications.